Lightweight distributed vibration sensing algorithm boosted with cross-domain distillation

We present an innovative real-time processing framework for long-distance phi-OTDR distributed vibration sensing (DVS) systems. This solution combines an ultra-lightweight convolutional neural network (CNN) with FPGA-based hardware acceleration and cross-domain distillation technique. By embedding prior physical knowledge into the model, the cross-domain distillation method significantly improves the efficiency and interpretability of the neural network models. The proposed system, based on a high-performance model containing just 4,141 parameters, achieves a processing speed of 0.019 ms per sample, enabling real-time analysis of raw DVS data from optical fibers extending up to 168.7 km. This research addresses the resource constraints in practical real-time DVS applications, enhancing the scalability and practicality of DVS systems while enabling their adaptability to broader, more diverse, and demanding application situations.